Running Gemma inside an office printer: 14MB RAM, 60s per token
TheTuringPost · x · 2026-11-18
osanseviero compiled a small transformer engine + quantized Gemma 3 270m into a single PostScript job with Carbon's help, running 100% inside an office printer: 14MB peak RAM, 60s per token, with layer juggling to avoid OOM. Top reply: "once the ink runs out, it'll refuse to generate the next token."
More from coding & agent
- The killer app of agentic programming is building harnesses for agentic programming — danbri · 2026-11-18
- Dev shows agentic video editing with HyperFrames + Claude, ditching FFmpeg pain — toolstelegraph · 2026-11-18
- ETH's AuXSpec: executable postconditions as RL reward, 27B model beats frontier LLMs at SpecGen — eth-algopt · 2026-11-18
- dottxt tool-call index: worst provider endpoint breaks 15x more tool calls than best — remilouf · 2026-11-18
- Dev runs Claude+Codex+Grok like a team, ships two releases in a day — GeneralDistinct5775 · 2026-11-18
- GAUSSIAN EVIL: a browser horror game homage to P.T. built with 3D Gaussian splatting — willeastcott · 2026-11-18